Land Cover Classification Using Remote Sensing Image Based on Support Vector Machines

Humans have altered land surface since ancient times,and land cover has a profound influence on global environmental change.The environmental changes(local or global) in turn affect humans.Since the 1990s,there has been more and more attention paid to the land use/cover issues that affect human society.Remote sensing is a useful tool for monitoring land cover change,and people can use the land cover maps or images to help make appropriate decisions regarding land use.We use classification of remote sensing images as a foundation for this study on land use/cover.The question of how to enhance the accuracy of image classification is a very important issue that has puzzled experts for many years.Traditional classification techniques such as maximum likelihood or isodata sometimes have a very low accuracy that restricts the results from being used as a reference for land use policy.This article uses the support vector machine(SVM) as a tool for classification of remote sensing images and compares it with traditional classification techniques such as maximum likelihood.The results indicated the SVM has a higher classification accuracy than maximum likelihood when using different parameter combinations.Its highest overall accuracy is higher than the maximum likelihood of 0.9779%.Although both techniques have a mixed-class phenomenon in the classification results,the mixing degree of SVM is lower than for maximum likelihood.Overall,the SVM classification results are in the scope which is acceptable for most applied work.Since SVM is based on statistical learning theory,its decision principle is structure risk minimization and VC dimension.When separating two classes,it guarantees the biggest interval between two classes,which means that the structure risk is minimal.Traditional classification techniques like maximum likelihood are based on mathematical statistics.This requires that remote sensing data has a normal distribution,but usually remote sensing data has a separate,multinomial distribution,so it is difficult for traditional techniques to achieve a high accuracy in image classification.SVM is a promising tool in the remote sensing image classification field.